Rebuilding a local voice assistant pipeline
vdf.ai

Rebuilding a local voice assistant pipeline

Archived — this story has rotated out of today’s deck. It is kept here in full.

The gist

Fabio Manganiello rebuilt Platypush's voice assistant pipeline, cutting post-speech lag from 6–10 seconds to about 1.6 seconds.
The fix targets the stalled conversations that plague local open-source voice setups.

Background

Platypush is a general-purpose Python platform for IoT and automation, and Fabio Manganiello is the author of its blog post describing the rebuild. After a couple of months using the previous pipeline as his main driver, he was unhappy with high false positive/negative rates and poor start/end-of-speech detection, especially in noisy rooms and with cheap microphones. The rebuild swaps openWakeWord — a project that has been quiet, with pre-trained v0.1 models from 2023 trained on synthetic speech — for microWakeWord, the wake word engine used by ESPHome and Home Assistant Voice.

How it unfolded

  1. Nov 19, 2026Manganiello publishes "Rebuilding my voice assistant pipeline," swapping openwakeword for microwakeword and reworking speech detection in Platypush.
  2. Nov 19, 2026Platypush adds the assistant.microwakeword plugin, fixes frame sizing and parameters in assistant.openwakeword, and reworks the shared AudioPreprocessor with multiple VAD backends, adaptive noise handling and a Silero VAD wrapper running ONNX inference through onnxruntime.
  3. Nov 19, 2026The rebuild cuts post-speech conversation lag from six to 10 seconds down to about 1.6 seconds.

Key terms

VAD
Detection of the presence or absence of human speech in an audio signal.

Source: Wikipedia

Who’s saying what

Author
Manganiello said the previous pipeline had high false positive/negative rates and poor start/end-of-speech detection, especially in noisy environments and with cheap microphones.
Analysts
LavX News said the VAD and noise-floor work addresses the biggest complaint about open-source pipelines — conversations that never close — while remaining limits are hardware and defaults tuned on synthetic audio.
Caution
LavX News noted the best latency still depends on the cloud: tts.piper lags three to four seconds on a Raspberry Pi 5 and assistant.openai needs network access.

Sources

See today’s stories in the app Gistline — free on the App Store